AI & software engineer · Machine LearningOrlando, FL, working remotely with teams anywherePortfolio ’26
Buildingintelligent software,end to end.
I'm Gui, a software engineer with a BSc in Computer Science specialising in AI & Machine Learning. I build applied AI end to end: RAG, agent workflows, prompt engineering, model evaluation and guardrails, shipped to the cloud with CI/CD, plus the APIs and interfaces people actually use.
My BSc specialism, and the thread running through my recent work. I like the whole pipeline: framing the problem, getting the data right, choosing the method, and shipping something people can use and trust. Since finishing my degree I've been self-studying Stanford CS231n (deep learning for computer vision).
Machine learning & data
Supervised learning, model evaluation and data analysis in Python, plus university projects in NLP and OCR.
PythonMLNLPOCR
Reinforcement learning
Q-learning, TD methods, DQN and policy-gradient methods like Proximal Policy Optimization (PPO), which powers my final project.
Q-learningDQNPPODeep RL
Evolutionary computation
Genetic encodings and selection, including evolving simulated creatures in a physics engine.
Genetic algorithmsPyBullet
Applied AI systems
RAG, agent-based workflows and orchestration, tool integration, prompt engineering, model evaluation and guardrails.
RAGLLM agentsEvalsGuardrails
ML lab: three tiny algorithms, running live in your browser
Supervised learning
Logistic regression, trained by gradient descent
Two noisy clusters. Each frame takes one gradient step on the log-loss, and the decision boundary rotates into place.
0epoch
–log-loss
–accuracy
Reinforcement learning
Value iteration in a grid world
Each sweep applies the Bellman update, and value spreads outward from the goal. Once it converges, the agent follows the greedy policy around the walls.
0sweeps
–max Δ
–path length
Evolutionary computation
A genetic algorithm solving OneMax
40 genomes of 48 bits each. Tournament selection, crossover and mutation push the population toward all ones. The chart plots best and mean fitness per generation.
0generation
21best / 48
14.9mean
Toy problems, computed live in the browser. They show the ideas I work with. They aren't results from my projects.
(02)
Featured project
BSc final project · University of London · AI & ML
A decision support system for ETF allocation.
It turns an investor's goals and risk profile into a recommended ETF allocation, and explains the recommendation.
To do this, I trained 80 reinforcement learning agents with four different rewards and tested them year by year from 2021 to 2024.
Rewards that penalise risk beat a plain return reward, mainly because they trade less. In the web app, the user picks the reward that matches their attitude to risk and gets the agent's allocation.
It is capped by fixed risk rules and comes with a full explanation of each agent's reasoning and recent news context.
A five-agent LLM pipeline explains the trade-offs in plain language, behind a guardrail: a test fails the build if any agent can touch the allocation. The user makes the final call.
Designed and run a self-hosted multi-agent AI platform on GCP: 9 LLM agents orchestrated through a gateway, with a FastAPI memory API and RAG over Qdrant and Neo4j, deployed with Docker.
Tool integration with GitHub, Telegram, Notion, OpenRouter and OpenAI, so agents automate tasks from request to pull request.
Guardrails: human approval gates, services bound to localhost and allowlisted access, so agents can't ship critical changes without review.
PromptFoo eval suites to regression-test prompts and agent skills before release, and tiered model routing through OpenRouter to cut costs.
Web & ops work for small-business clients.
2022 — 2024
CXP Brasil Consulting
Front-End Developer · Unimed
Health insurance
Built Vue.js interfaces for Unimed, Brazil's largest medical cooperative (24M+ members), raising patient portal engagement by 25%.
Created a reusable component library with Vue.js and React, cutting development time for new features by 30%.
Worked directly with client stakeholders in Agile/Scrum sprints.
2020 — 2022
Apprenticeship
Web Developer · Brussels
Agency
Built Vue.js front ends and Node.js/Express REST APIs with MySQL and MongoDB for 10K+ daily users.
Owned web apps from concept to deploy.
(05)
Toolkit & study
Applied AI
RAG
Agent workflows & orchestration
Prompt engineering
Model evaluation · PromptFoo
Guardrails · observability
ML & data
Machine learning
Deep RL (PPO)
NLP · OCR
LangChain · Hugging Face
Qdrant · Neo4j
Languages & APIs
Python · SQL
TypeScript · JavaScript
FastAPI · Node.js
RESTful APIs · PostgreSQL
Vue.js / Nuxt · React
Cloud & DevOps
GCP · serverless
Docker · Linux
GitHub Actions · CI/CD
Azure DevOps
Git
Practices
Agile / Scrum
Unit & contract tests
Jest · Vitest
Secure coding
Responsible AI
Education
BSc Computer Science University of London · AI & ML · 2026
Stanford CS231n Self-study · deep learning for computer vision
